Mentatcurated
Biotech & Synthetic Biology medium · first-party

The cell model that fought back

The AI models meant to predict how a cell reacts to switching a gene on or off keep losing to a plain straight line, and Xaira Therapeutics argues the fix is feeding them cause instead of correlation.

A month-old embarrassment in computational biology is that billion-parameter models trained to predict a cell's response to genetic tinkering often can't beat a regression line a statistician could have drawn in the 1970s. Xaira Therapeutics, a drug-discovery startup, has a diagnosis and a counterexample. On the Latent Space podcast its researchers argue the problem was never the model. It was the diet.

Bo Wang built scGPT — one of the observational models now accused of hitting a wall — and is now arguing against his own generation of work.

The usual training data is a giant catalog of cells sitting in their natural states, a snapshot of what genes happen to be active together. That is correlation. What a model needs to predict the effect of *changing* a gene is causation, and the only way to get that is to actually flip genes and watch. Xaira built X-Atlas/Pisces, a library of roughly 25 million cells whose genes were silenced one at a time with CRISPR and then read out. Trained on that, its X-Cell model's accuracy keeps climbing as the model grows, following the same smooth improvement curve that made large language models work, where the observational data flatlines. Xaira reports the causal data carries about thirty times more usable signal, and that X-Cell finally clears the linear baseline.

The pointed wrinkle: Bo Wang, who leads the AI work, built scGPT, one of the influential observational models now accused of hitting a wall. He is arguing against his own prior generation of work. The claims are worth watching rather than banking, all of them are self-reported in a preprint from a well-funded startup, unreplicated and not peer reviewed, and one rival founder notes that predicting a lone cell's behavior is still a long way from predicting a patient's. But the released model and a slice of the dataset are out, which is the kind of homework the rest of the field can check.

The lenses

Novelty 3
Impact · breadth 2
Impact · depth 3
Actionable 3
Delight 3
Substance 4
Hype 2

The facts

What it isA preprint plus a released model and a public slice of its dataset, not a productized tool
The claimFeeding a cell model CRISPR-perturbation data restores LLM-style scaling where observational data plateaus
StatusSelf-reported by the maker, not yet independently replicated or peer reviewed
The wrinkleIts AI lead built an earlier observational model he is now arguing against
Open latent.space →

How this connects

Tap a node to open it

How this connects